Attention Mechanism-Based CNN-LSTM for Abusive Comments Detection and Classification in Social Media Text
BalaAnand Muthu First, Kogilavani Shanmugavadive, Sathishkumar Veerappampalayam Easwaramoorthy, Muthukumaran Maruthappa, Malliga Subramanian, Rajermani Thinakaran · International Journal of Advanced Computer Science and Applications · 2024
Human contact with one another through social networks, blogs, forums, and online news portals and communication has dramatically increased in recent years. People use these platforms to express their feelings, but sometimes hateful comments are also spread. When abusive language is used in online comments to attack individuals such as celebrities, politicians, and products, as well as groups of people associated with a given country, age, or religion, cyberbullying begins. Due to the ever-growing number of messages, it is challenging to manually recognize these abusive comments on social media platforms. This research work concentrates on a novel attention mechanism-based hybrid Convolutional Neural Network - Long Short Term Memory (CNN-LSTM) model to detect abusive comments by getting more contextual information from individual sentences. The proposed attention mechanism-based hybrid CNN-LSTM model is compared with various models on the dataset provided by the shared task on Abusive Comment Detection in Tamil – ACL 2022 which contains 9 class labels such as Misandry, Counter-speech, Xenophobia, Misogyny, Hope-speech, Homophobia, Transphobic, Not-Tamil and None-of-the-above. We obtained an accuracy of 67.14%, 68.92%, 65.35% and 68.75% on Naïve Bayes, Support Vector Machine, Logistic Regression and Random Forest respectively. Furthermore, we applied the same dataset to deep learning models like Convolutional Neural Networks (CNN), Long Short Term Memory (LSTM), Bidirectional-Long Short Term Memory (Bi-LSTM) and obtained an accuracy of 70.28%, 71.67% and 69.45%, respectively. To obtain more contextual information semantically a novel attention mechanism is applied to the hybrid CNN-LSTM model and obtained an accuracy of 75.98% which is an improvement over all the developed models as a process innovation.